AML/CFT pattern detection and the privacy tradeoff

Three ways to see it

1. 1

   AML AI in banking does pattern detection at three levels. Account-level: structuring, unusually round amounts, sudden activity in dormant accounts. Network-level: cycles of transfers, mule accounts, related parties. Cross-institution level: same actor moving across banks. Pakistani banks have variable capacity across these levels. The first is broadly automated. The second is partly. The third is mostly manual and depends on STR sharing through the FMU. AI helps most at the second level today.

2. 2

   Way one to think about the privacy tradeoff: pseudonymise for analysis, re-identify on hit. The model can run on transaction patterns without ever seeing customer names. Only when a pattern crosses a threshold does the compliance officer get the names. This is more privacy-preserving than the typical Pakistani bank's current arrangement, where the analyst sees customer names from the first dashboard click.

3. 3

   Way two: false positive economics. Every alert costs an officer 10 to 30 minutes of work. A bank with 5,000 alerts a day at five percent precision is burning a hundred officer-hours daily on noise. AI that pre-classifies alerts, suppresses obvious false positives, and concentrates officer attention on the top 5 percent of priority signals is the workforce multiplier the AML function needs. The privacy gain is collateral; fewer officers reading more relevant records.

The why-tree

Try this with Claude

AI-edge prompt to try: 'Acting as an FMU-aligned compliance head, propose ten AI-detectable AML patterns active in Pakistan, the data each requires, and one false positive class I should expect.' Use as a brainstorm.